A Design Of Massive MIMO Detectors Via Deep Neural Networks

Recently,in wireless communication systems,the massive multi-input multi-output(MIMO)technology has attracted significant attention in both the industry and academia sectors.In the fifth generation(5G)mobile networks and beyond,the data traffic is expected to increase exponentially.The massive MIMO technology can serve such systems without overuse of the scarce resources,i.e.,without increasing the transmit power or allocating more bandwidth.In an ideal configuration,massive MIMO is equipped with an unlimited number of base station(BS)antennas to serve many tens of user terminals(UTs)in the same time-frequency resource.Though the massive MIMO can handle large data sizes,the conventional signal processing algorithms have inherent limitations to satisfy the high-speed,large data requirement of this system.As a result,the massive MIMO requires highly efficient signal processing algorithms to serve its promising benefits.In particular,since the computational complexity forbids the implementation of the classical MIMO detectors,an efficient detection scheme is highly required.On the other hand,deep learning(DL)has become the most successful machine learning method in various applications,such as computer vision and natural language processing.Due to their ability to learn a proper hierarchical representation of data,the DL can solve more complex problems where rigid mathematical models cannot be used.Besides the above domains,the DL has shown promising results in communication systems,including the physical layer.MIMO detection is one of those applications in which the DL has attracted considerable attention.Most of the recent works employ the deep unfolding strategy to transform an existing algorithm to the DL layers and improve the performance through learning.However,as it is a new field of application,there are significant limitations to these works.Either their performance is not closer to the optimum detector,or their computation is costly.As such,this thesis investigates the application of two broad categories of DL approaches,data-driven and model-driven,for the massive MIMO data detection.Firstly,we study the data-driven approach;and then design deep neural network(DNN)detectors that can reduce the computational complexity with comparable performances.Secondly,we investigate the model-driven approaches;and apply the popular deep unfolding method to design an efficient detection framework that can improve the performances of the state-ofthe-art detectors.More precisely,the main features of this thesis are summarized as follows:(1)The generic DNN architectures can not efficiently learn to detect under the randomly varying channel scenario when directly working with the noisy received signal.In this work,we present a design of DNN detectors that can overcome this challenge.First,we explore some pre-processing techniques to extract relevant features.Then we formulate the MIMO detection problem as a set of independent binary classification tasks by adopting a binary relevance algorithm that decomposes a multi-label learning task into a set of independent binary classification tasks.Using this representation,we design customized DNN detectors that can efficiently detect transmitted symbols under the classical and randomly varying independent and identically distributed(i.i.d.)Rayleigh channel scenario.Besides,we also extend this approach to spatially correlated channels by using an exponential correlation matrix.Our method is simple and robust against the randomly varying channels and additive noise;i.e.,once trained offline,the networks can detect efficiently under the same distribution without requiring new training.Numerical results show that the proposed detectors can achieve better symbol error rate performances with similar or cheaper computational costs compared to several existing traditional and DNN-based detectors.(2)The optimum detector is impractical for the massive MIMO system since the computational complexity exponentially increases with the problem size.The alternating direction method of multipliers(ADMM)-based detectors can achieve sub-optimum performance.However,they have two main limitations:First,it is challenging to choose the penalty parameters.Second,their computational complexity is higher compared to the classical linear detectors.In this work,we propose two deep neural network(DNN)-based detectors that can overcome these limitations:(1)We design a DNN architecture that can efficiently approximate the adaptive penalty parameters.(2)We propose a low-complexity detection framework that can achieve a better performance when the number of receiving antennas is much larger than the number of transmitting antennas.Further,we introduce a feature selection procedure that can lead to efficient learning in a method like this.Numerical results show that the proposed method can achieve better performance with equal or less computational complexities than the state-of-the-art detectors.